Accelerometer data is a stream of numbers representing acceleration forces along three axes (X, Y, and Z), and interpreting it means separating gravity from movement, filtering out noise, and translating the remaining signal into something meaningful like steps, activity intensity, or sleep. The process involves several layers, from understanding the raw coordinate system to choosing the right settings for your analysis. Whether you’re working with a research-grade device, a smartphone, or a consumer wearable, the core principles are the same.
What Raw Accelerometer Data Looks Like
An accelerometer measures acceleration in three perpendicular directions. When the device is sitting still on a table, you won’t see all zeros. You’ll see approximately 9.81 m/s² (1g) on whichever axis points toward the ground, because the sensor constantly detects gravitational pull. The other two axes will read close to zero. If the device is tilted, gravity gets split across multiple axes. This is the first thing to understand: your raw data always contains gravity mixed in with actual movement.
To isolate dynamic motion (walking, arm swings, vibrations), you need to subtract or filter out the gravitational component. One common approach is to have the user stand still for about 10 seconds, average those readings to establish a gravity vector, and then mathematically rotate all subsequent data so gravity falls on a single known axis. A rotation matrix corrects for whatever angle the device happens to be tilted at. After this correction, any acceleration on the other axes represents real movement rather than a quirk of how the sensor was oriented.
If you skip this step, a phone in a tilted pocket will show constant acceleration on axes that should be quiet, and your movement estimates will be off. Many consumer devices and apps handle this automatically, but if you’re working with raw data yourself, orientation correction is essential before any further analysis.
Filtering Noise From the Signal
Raw accelerometer signals contain a lot of information you don’t want. Low-frequency drift comes from slow postural changes. High-frequency spikes come from vibrations, vehicle travel, or electrical noise in the sensor itself. The useful signal for human movement sits in a specific frequency band.
Acceleration related to human locomotion falls between 1 and 10 Hz. Walking produces rhythmic signals around 1 to 3 Hz, while running and vigorous movements push higher. Noise from sources like car driving tends to show up above 10 Hz. A well-chosen filter keeps the movement signal and discards the rest.
The most common approach is a bandpass filter, which cuts frequencies below a low threshold and above a high threshold. The original ActiGraph filter (widely used in physical activity research) has cutoff frequencies at 0.29 Hz on the low end and 1.63 Hz on the high end. That high-end cutoff is quite aggressive, though, and can trim out vigorous activity. More recent recommendations suggest keeping everything up to 10 Hz as the optimal low-pass cutoff for capturing the full range of human movement. A fourth-order Butterworth filter is a standard choice that provides a smooth frequency response without distorting the signal shape.
The high-pass cutoff at around 0.29 Hz allows signals with cycles as slow as about 3 seconds to pass through, which covers slow walking. Setting it lower risks letting gravitational drift leak back in. Setting it higher risks missing gentle movements.
Epoch Length Changes Your Results
After filtering, accelerometer data is typically summarized into “epochs,” which are time windows that condense the continuous signal into a single value per window. Common epoch lengths range from 1 second to 60 seconds. This choice has a surprisingly large effect on your results.
Longer epochs average out short bursts of activity. A 10-second sprint followed by 50 seconds of sitting looks very different in a 1-second epoch (where the sprint is clearly visible) versus a 60-second epoch (where it gets diluted into a low average). Research comparing epoch lengths found that shifting from 1-second to 60-second epochs decreased reported sedentary time by up to 28.7%. That’s not a small difference. Longer epochs also reduce time classified as moderate or vigorous activity while inflating time classified as light activity.
For research or clinical purposes, shorter epochs (1 to 5 seconds) capture intermittent activity patterns more accurately. For general daily summaries where you just want overall activity levels, longer epochs can be acceptable. The key point: if you’re comparing data across different sources or studies, the epoch length must match, or the numbers won’t be comparable.
Translating Counts Into Activity Intensity
Most accelerometer software converts filtered, epoch-summarized data into “counts,” an arbitrary unit representing the total acceleration detected in each time window. These counts are then classified into intensity categories using established thresholds called cut-points.
The most widely used cut-points come from a set developed by Troiano and colleagues, which define moderate-to-vigorous physical activity (MVPA) as anything above 2,020 counts per minute. Below that, you’re in light activity or sedentary territory. These thresholds were derived by having people walk and run on treadmills while measuring their actual energy expenditure.
However, these absolute cut-points don’t account for individual fitness levels. A highly fit person might find 2,020 counts per minute barely noticeable, while someone less fit might already be working hard. Individualized cut-points based on perceived effort or heart rate tend to be higher: around 4,004 counts per minute when based on perceived exertion, and around 3,187 counts per minute when based on heart rate. These also vary by sex, with female-specific thresholds running somewhat lower (2,951 counts per minute by heart rate) compared to male-specific ones (3,503 counts per minute).
If you’re using a consumer device, this translation happens behind the scenes and typically uses the manufacturer’s own algorithm. If you’re analyzing raw data yourself, choosing which cut-points to apply is one of the most consequential decisions in your analysis.
Where You Wear It Matters
Device placement on the body significantly affects accuracy, particularly for step counting. A hip-worn accelerometer using default settings showed only 12% accuracy for step detection at slow walking speeds (below 0.6 m/s, which is a very slow shuffle). At normal walking speeds above 1.0 m/s, accuracy jumped above 92%. An ankle-worn device performed much better at slow speeds, reaching 71% accuracy with default settings and 96% with a low-frequency extension filter applied.
The pattern holds during turning movements too. Hip-worn devices with default filters detected only 25% of steps during slow walking with continuous turns, while ankle-worn devices caught 82% with default settings and 96% with the extended filter. At moderate and fast walking speeds, both placements perform well regardless of filter choice.
The practical takeaway: if you’re tracking activity in older adults or anyone who moves slowly, ankle placement is more reliable. Hip placement works well for moderate to brisk walking. Wrist-worn devices (the most common consumer placement) introduce their own complications because arm movements during non-walking activities can register as steps, while keeping your hands still on a shopping cart can cause missed steps.
Interpreting Sleep Data From Accelerometers
Accelerometers detect sleep by identifying sustained periods of immobility. The device doesn’t measure brain waves or sleep stages directly. Instead, scoring algorithms look for consecutive minutes where the activity count drops to zero or near zero. Common rules define sleep onset as the first minute of 3, 5, 10, 15, or 20 consecutive immobile minutes. For the longer windows (10, 15, and 20 minutes), one minute of movement is typically allowed within the period, since brief position shifts during sleep are normal.
Two algorithms dominate sleep research: the Sadeh algorithm and the Cole-Kripke algorithm. Both analyze a rolling window of activity counts around each minute to classify it as sleep or wake. They perform reasonably well at detecting total sleep time but tend to overestimate sleep because they classify any quiet wakefulness (lying still in bed reading, for example) as sleep. They also struggle with sleep onset latency, the time it takes to fall asleep, because people often lie relatively still before actually sleeping.
Features for Activity Classification
Beyond simple intensity categories, accelerometer data can be used to identify specific activities like walking, cycling, or climbing stairs. This requires extracting features from the signal and feeding them into a classification model. The most commonly used features in time-domain analysis are the mean acceleration value, the variance (how much the signal fluctuates), and skewness (whether fluctuations lean toward higher or lower values). Frequency-domain features like spectral entropy capture how rhythmic or chaotic the movement pattern is.
Walking produces a highly rhythmic signal with clear peaks at regular intervals. Cycling shows a different frequency pattern with less vertical bounce. Sitting in a car produces high-frequency vibration above 10 Hz that looks nothing like human movement once you know what to look for. Machine learning models trained on labeled examples of each activity can classify new data with high accuracy, but the quality of the input features determines the ceiling of performance.
Common Pitfalls in Interpretation
Several issues can lead to misleading results if you’re not watching for them. Non-wear time is one: if the device was removed and sat on a table, that period looks identical to sedentary behavior. Most analysis protocols require at least 10 hours of valid wear time per day and flag strings of consecutive zeros lasting 60 minutes or more as likely non-wear periods.
Comparing data between different devices is another frequent problem. Each manufacturer uses different sensors, filters, epoch settings, and proprietary algorithms. Counts from one brand are not equivalent to counts from another, even if both devices were worn simultaneously. Validation studies using doubly labeled water (the gold standard for measuring energy expenditure) show that some devices achieve intraclass correlation coefficients above 0.80 in free-living conditions, but this level of agreement is not universal across all devices or populations. Accuracy tends to be lower in overweight and obese individuals, where movement patterns and energy costs differ from the populations used to develop most algorithms.
Water-based activities, cycling, and upper-body exercises are poorly captured by hip or wrist accelerometers because they don’t produce the vertical oscillation these devices are optimized to detect. If these activities make up a significant portion of someone’s routine, accelerometry alone will underestimate their total physical activity.